[ExI] DeepMind wins third Go game and the championship

[Adrian Tymes]

On Sat, Mar 12, 2016 at 9:45 AM, William Flynn Wallace <foozler83 at gmail.com>
wrote:

> Is the art of bluffing programmable?
>

Yes.

You have to have a good enough model of the ones you're bluffing - but
that's exactly like humans needing to know something about the ones they're
bluffing: someone from a completely different culture and background, who
you have no information about and no way to read, you don't even know where
to begin.  Same with an AI bluffer.

Once you have that model, you run that model with the information you know
about, to assess the chances that the other side will call your bluff.  You
then compare to the advantage you would gain from bluffing, and do a
risk/reward computation.  Again, this is exactly what human bluffers do
too, though they may do it less formally.

The difficult part is forming those models.  This is, at some level,
understanding one's opponent - and being able to understand another person
at this level is one of the traits that humans prize as making themselves
humans.  Those humans who utterly lack this capability at any meaningful
level (not necessarily enough to win poker, but enough to usefully
anticipate others' reactions to their actions) are viewed as damaged, in
need of repair if possible.

This might not be everything that's needed for "true" AI, but a generalized
capability to create these models would certainly be one step closer.  And
once programs can do this at near-human levels, that will immediately have
utility on its own: at first to teach the "damaged" ("sociopathic") humans
what they are missing and how to (re)gain this ability that is necessary to
function in society, then (as the capability improves) in automating
certain standard social interaction tasks, such as sales.

One could imagine a future where there exist salesbots and anti-salesbots,
the former attempting to negotiate an exchange of benefit to the former's
users which may or may not be of benefit to the latter's users, the latter
attempting to screen out those not of benefit to their users.  One could
even freak out and hypothesize that the higher version will always win -
that each moment-by-moment incremental improvement will somehow translate
into an inevitable victory for the latest version, which would more likely
be used by those with more resources, and that since this affects said
resources that this would inevitably snowball into an extreme parody of
today's concentrations of wealth.  (Most likely, quite a few people will
have this exact freak out once they realize salesbots are theoretically
eventually possible - among other factors, that it is an extreme version of
a widely recognized modern social problem makes it attractive - which makes
specific countermemes worth preparing.)

But that seems unlikely to happen in practice.  For one, those who have few
resources can not make expensive transactions by definition - you can't
squeeze blood from a stone - which inherently limits the value that the
salesbots could extract.  For another, that even the wealthy suffer when
most people are in poverty have been proven: to be moderately wealthy in an
affluent society is, in most modern and historical examples, to be
wealthier in absolute terms than to be the richest person among the
desperately poor.  For a third - correlated with the second - many things
are only affordable if purchased en masse, and get better the more people
they serve.  (One example: computer chips.  It would be basically
impossible for even the richest man on Earth to fund a modern fab
development program for his exclusive use, but when that same fab makes
chips with millions or billions of customers, there is a better than
billion-fold improvement.)  This information would be available to these
hypothetical super-sales agents, who - if they were tasked with maximizing
their owners' wealth, as may well happen - would go about making everyone
rich (by the standards around when they started) because that would be the
most efficient method of accomplishing their goal (as is said, a rising
tide lifts all boats).

Then there's the matter of incremental improvement being inevitably that
much better.  To be blunt, that just doesn't happen in reality, in any
field.  There is no reason to believe that salesbot version 5.4.8.4 will
automatically completely defeat all anti-salesbots of version 5.4.8.3 and
earlier.  Indeed, there have been times where later versions were actually
worse; one current example is the large number of Windows 7 users who have
judged Windows 8 and 10 to be downgrades (while their judgement is
subjective, one can suppose that a sizable majority of competent users will
recognize and act in their own self interest, and that - absent specific
evidence to the contrary - such a majority is more likely than not to be
correct), to the point that Microsoft is having to support Windows 7 more
than anticipated.

Also, these are voluntary transactions.  If the salesbot comes from a
sufficiently distrusted source - like, say, a certain wealthy individual
known or believed to have acquired that wealth through less than mutually
beneficial transactions - that salesbot may simply run into a wall of
everyone saying "no", to which the salesbot has no recourse.  (At least,
none save for resorting to misrepresentations and other actions commonly
deemed damaging to the common good and therefore illegal in most places -
but if we're supposing super salesbots, let us also suppose merely
competent fraud analytics of the sort some law enforcement agencies already
have access to, which would be able to detect these activities, and then
the super salesbots would find themselves deactivated when their users are
jailed.  The law could be bought off, but that only works temporarily:
eventually someone gets into power who hasn't been bought off.
Sufficiently smart salesbots would know this would be the eventual outcome
if they were abusive, and thus that it would be in service to their end
goal to not be abusive.)

Is that a sufficiently long answer?  ;)
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20160312/e1d55e57/attachment.html>